Abstract

This thesis tackles the obstacles of object detection, tracking, and pose estimation in augmented reality (AR) applications by developing a comprehensive system using synthetic datasets. To achieve this, the research utilizes YOLOv7 and YOLOv8 for object detection, Deep SORT for object tracking, and SC6D and YOLO-Pose for 6DoF (Six Degrees of Freedom) pose estimation. To overcome the limitations of obtaining accurate ground truth poses from real datasets, a synthetic dataset is created. This dataset contains a diverse range of virtual objects in various poses, lighting conditions, and backgrounds. It serves as a representative sample to train and evaluate the system components’ performance. The first part of the research focuses on training and assessing the effectiveness of YOLOv7 and YOLOv8 for object detection using the synthetic dataset. The models are fine-tuned on synthetic images to accurately detect objects in real-time. The evaluation includes precision, recall, and average precision metrics to determine the models’ effectiveness in object detection for the AR system. The second part concentrates on evaluating 6DoF pose estimation using SC6D and YOLO- Pose, trained on the synthetic dataset. The pose estimation methods are evaluated by measuring translation and rotation errors to assess their accuracy and robustness in estimating object poses for AR applications. The results demonstrate that using synthetic datasets for training and evaluating object detection, tracking, and pose estimation methods in AR systems is effective. By leveraging synthetic data, this thesis contributes to overcoming the challenges of obtaining accurate ground truth poses from real datasets.

Results/Implementation/Project Description

Conclusion

[ PDF (optional) ] 

[ Slides Kickoff/Final (optional)]